Lec 01 - Deep Learning for Computer Vision

https://www.youtube.com/playlist?list=PL5-TkQAfAZFbzxjBHtzdVCWE0Zbhomg7r

(첨부된 이미지 및 강의 내용 출처)

Deep Learning for Computer Vision

 

Summary

what is AI?
making machines do things that people normally do

cf) there are types of AI which have nothing to do with "learning" or "deep learning" like symbolic systems

what is CV?
1 teaching machines to see and machine to learn
2 Building artificial systems that process, perceive, and reason about visual data

   examples)
    1 autonomous vehicles augmented
    2 virtual reality drones

what is the machine learning?
the process of building artificial systems that learn from data and experience

what is the deep learning?
1 another subset of machine learning
2 Hierarchical learning algorithms with many “layers”, (very) loosely inspired by the brain

our target :

 

                                                   • A brief history of computer vision 

1. Hubel and Wiesel, 1959
        they wanted to understand how the mammalian brains work
            electrode into cat brain(visual cortex)
            cat watching different sorts of slides
            record the neural activity

    hypothesis : there's certain neurons in the brain that responds different types of visual stimuli

 

    discovery :

     what kinds of images would activate the neurons?

     there are differnet types of cells in the brain that are responding to different types of visual stimuli

     1 Simple cells : response to light orientation
                          respond to and edge light on one side dark on another side

     2 Complex cells : response to light orientation and movement

     3 Hypercomplex cells : response to movement with an end point

     this stusy is beginning of computer vision
        reason1 : this emphasis on oriented edges 
        reason2 : this shows hierarchical representation of the visual system of building from simple cells, complex                           cells, and more more complex cells


2. Larry Roberts, 1963 
    thesis : how do you actually get photographic information into the computer?
    >> detect some of the edges in the picture -inspired by wiesels(edges were fundamental to visual processing)

 

3. David Marr 1970s

 

4. recognition via parts 1970s
idea : generalized cylinders, pictorial structures
(but this era was AI winter)

5. recognition via edges  1980s
John canny, 1986
   detecting edges, matching edges

6. recognition via grouping 1990s
image segmented into semantically meaningful chunks

7. recognition via matching 2000s
SIFT by David lowe, 1999
    finding invariant robust feature in image,
    matching correspond points in the one image into points in the other image

8. Viola and Jones, 2001
face detection
    one of the first successful applications of machine learning to vision
    it was the first major use of machine learning and CV

 

why is it important?
   1 boosted decision trees algorithm 
   2 very fast commercialization -shipped in digital camera

                                           (boxes are on the faces and focus on the people in the scecne)

9. PASCAL visual object challenge, 2001

10. IMAGENET large scale visual recognition classification challenge

2012, finally deep learning break through CV
AlexNet :Krizhevsky, Sutskever, and Hinton, NeurIPS 2012

 

                                                            • A brief history of deep learning

1. perceptron, 1958, by frank rosenblatt
Could learn to recognize letters of the alphabet from data 

2. Minsky and Papert, 1969
Showed that Perceptrons could not learn the XOR function

3. Neocognitron: Fukushima, 1980
Computational model the visual system, directly inspired by Hubel and Wiesel’s hierarchy of complex and simple cells  Interleaved simple cells (convolution) and complex cells (pooling)

his model Looks a lot like AlexNet 
but he doesn't have practical training algoritm

4. Backpropagation: Rumelhart, Hinton, and Williams, 1986
backprop algorithm was Successfully trained  perceptrons with multiple  layers

5. Convolutional Networks: LeCun et al, 1998
it looks very much like fukushima algorithm
Applied backprop algorithm to a Neocognitron-like architecture

6. 2000s: “Deep Learning”
Not a mainstream research topic at this time
  Hinton and Salakhutdinov, 2006
  Bengio et al, 2007
  Lee et al, 2009
  Glorot and Bengio, 2010

7. 2012 to Present: Deep Learning Explosion
ConvNets are everywhere
  Image Classification
  Image Retrieval
  Object Detection - Ren, He, Girshick, and Sun, 2015 <<< I'm intersted in 
  Image Segmentation
  Video Classification - Simonyan et al, 2014 <<< I'm intersted in
  Activity Recognition <<< I'm intersted in
  Pose Recognition
  Playing Atari games
  Medical Imaging
  Whale recognition
  Galaxy Classification
  Image Captioning - Vinyals et al, 2015 Karpathy and Fei-Fei, 2015

Deep Learning Explosion was a combination of three big components algorithms, data, and computation

남은 과제 :
사진, 비디오를 보고 상황, 맥락을 이해하는 수준의 CV by DeepLearning